SageBwd: A Trainable Low-bit Attention

Low-bit attention, such as SageAttention, has emerged as an effective approach for accelerating model inference, but its applicability to training remains poorly understood. In prior work, we introduced SageBwd, a trainable INT8 attention that quantizes six of seven attention matrix multiplications while preserving fine-tuning performance. However, SageBwd exhibited a persistent performance gap to full-precision attention (FPA) during pre-training. In this work, we investigate why this gap occurs and demonstrate that SageBwd matches full-precision attention during pretraining. Through experiments and theoretical analysis, we reach a few important insights and conclusions: (i) QK-norm is necessary for stable training at large tokens per step, (ii) quantization errors primarily arise from the backward-pass score gradient dS, (iii) reducing tokens per step enables SageBwd to match FPA performance in pre-training, and (iv) K-smoothing remains essential for training stability, while Q-smoothing provides limited benefit during pre-training.

SLA2: Sparse-Linear Attention with Learnable Routing and QAT

Sparse-Linear Attention (SLA) combines sparse and linear attention to accelerate diffusion models and has shown strong performance in video generation. However, (i) SLA relies on a heuristic split that assigns computations to the sparse or linear branch based on attention-weight magnitude, which can be suboptimal. Additionally, (ii) after formally analyzing the attention error in SLA, we identify a mismatch between SLA and a direct decomposition into sparse and linear attention. We propose SLA2, which introduces (I) a learnable router that dynamically selects whether each attention computation should use sparse or linear attention, (II) a more faithful and direct sparse-linear attention formulation that uses a learnable ratio to combine the sparse and linear attention branches, and (III) a sparse + low-bit attention design, where low-bit attention is introduced via quantization-aware fine-tuning to reduce quantization error. Experiments show that on video diffusion models, SLA2 can achieve 97% attention sparsity and deliver an 18.6x attention speedup while preserving generation quality.

LuSeeL: Language-queried Binaural Universal Sound Event Extraction and Localization

Most universal sound extraction algorithms focus on isolating a target sound event from single-channel audio mixtures. However, the real world is three-dimensional, and binaural audio, which mimics human hearing, can capture richer spatial information, including sound source location. This spatial context is crucial for understanding and modeling complex auditory scenes, as it inherently informs sound detection and extraction. In this work, we propose a language-driven universal sound extraction network that isolates text-described sound events from binaural mixtures by effectively leveraging the spatial cues present in binaural signals. Additionally, we jointly predict the direction of arrival (DoA) of the target sound using spatial features from the extraction network. This dual-task approach exploits complementary location information to improve extraction performance while enabling accurate DoA estimation. Experimental results on the in-the-wild AudioCaps dataset show that our proposed LuSeeL model significantly outperforms single-channel and uni-task baselines.

E2E-AEC: Implementing an end-to-end neural network learning approach for acoustic echo cancellation

We propose a novel neural network-based end-to-end acoustic echo cancellation (E2E-AEC) method capable of streaming inference, which operates effectively without reliance on traditional linear AEC (LAEC) techniques and time delay estimation. Our approach includes several key strategies: First, we introduce and refine progressive learning to gradually enhance echo suppression. Second, our model employs knowledge transfer by initializing with a pre-trained LAECbased model, harnessing the insights gained from LAEC training. Third, we optimize the attention mechanism with a loss function applied on attention weights to achieve precise time alignment between the reference and microphone signals. Lastly, we incorporate voice activity detection to enhance speech quality and improve echo removal by masking the network output when near-end speech is absent. The effectiveness of our approach is validated through experiments conducted on public datasets.

FlowSE-GRPO: Training Flow Matching Speech Enhancement via Online Reinforcement Learning

Generative speech enhancement offers a promising alternative to traditional discriminative methods by modeling the distribution of clean speech conditioned on noisy inputs. Post-training alignment via reinforcement learning (RL) effectively aligns generative models with human preferences and downstream metrics in domains such as natural language processing, but its use in speech enhancement remains limited, especially for online RL. Prior work explores offline methods like Direct Preference Optimization (DPO); online methods such as Group Relative Policy Optimization (GRPO) remain largely uninvestigated. In this paper, we present the first successful integration of online GRPO into a flow-matching speech enhancement framework, enabling efficient post-training alignment to perceptual and task-oriented metrics with few update steps. Unlike prior GRPO work on Large Language Models, we adapt the algorithm to the continuous, time-series nature of speech and to the dynamics of flow-matching generative models. We show that optimizing a single reward yields rapid metric gains but often induces reward hacking that degrades audio fidelity despite higher scores. To mitigate this, we propose a multi-metric reward optimization strategy that balances competing objectives, substantially reducing overfitting and improving overall performance. Our experiments validate online GRPO for speech enhancement and provide practical guidance for RL-based post-training of generative audio models.

TurboDiffusion: Accelerating Video Diffusion Models by 100-200 Times

We introduce TurboDiffusion, a video generation acceleration framework that can speed up end-to-end diffusion generation by 100-200x while maintaining video quality. TurboDiffusion mainly relies on several components for acceleration: (1) Attention acceleration: TurboDiffusion uses low-bit SageAttention and trainable Sparse-Linear Attention (SLA) to speed up attention computation. (2) Step distillation: TurboDiffusion adopts rCM for efficient step distillation. (3) W8A8 quantization: TurboDiffusion quantizes model parameters and activations to 8 bits to accelerate linear layers and compress the model. In addition, TurboDiffusion incorporates several other engineering optimizations. We conduct experiments on the Wan2.2-I2V-14B-720P, Wan2.1-T2V-1.3B-480P, Wan2.1-T2V-14B-720P, and Wan2.1-T2V-14B-480P models. Experimental results show that TurboDiffusion achieves 100-200x speedup for video generation even on a single RTX 5090 GPU, while maintaining comparable video quality. The GitHub repository, which includes model checkpoints and easy-to-use code, is available at https://github.com/thu-ml/TurboDiffusion.

MELA-TTS: Joint transformer-diffusion model with representation alignment for speech synthesis

This work introduces MELA-TTS, a novel joint transformer-diffusion framework for end-to-end text-to-speech synthesis. By autoregressively generating continuous mel-spectrogram frames from linguistic and speaker conditions, our architecture eliminates the need for speech tokenization and multi-stage processing pipelines. To address the inherent difficulties of modeling continuous features, we propose a representation alignment module that aligns output representations of the transformer decoder with semantic embeddings from a pretrained ASR encoder during training. This mechanism not only speeds up training convergence, but also enhances cross-modal coherence between the textual and acoustic domains. Comprehensive experiments demonstrate that MELA-TTS achieves state-of-the-art performance across multiple evaluation metrics while maintaining robust zero-shot voice cloning capabilities, in both offline and streaming synthesis modes. Our results establish a new benchmark for continuous feature generation approaches in TTS, offering a compelling alternative to discrete-token-based paradigms.

FLASepformer: Efficient Speech Separation with Gated Focused Linear Attention Transformer

Speech separation always faces the challenge of handling prolonged time sequences. Past methods try to reduce sequence lengths and use the Transformer to capture global information. However, due to the quadratic time complexity of the attention module, memory usage and inference time still increase significantly with longer segments. To tackle this, we introduce Focused Linear Attention and build FLASepformer with linear complexity for efficient speech separation. Inspired by SepReformer and TF-Locoformer, we have two variants: FLA-SepReformer and FLA-TFLocoformer. We also add a new Gated module to improve performance further. Experimental results on various datasets show that FLASepformer matches state-of-the-art performance with less memory consumption and faster inference. FLA-SepReformer-T/B/L increases speed by 2.29x, 1.91x, and 1.49x, with 15.8%, 20.9%, and 31.9% GPU memory usage, proving our model's effectiveness.

Exploring Efficient Directional and Distance Cues for Regional Speech Separation

In this paper, we introduce a neural network-based method for regional speech separation using a microphone array. This approach leverages novel spatial cues to extract the sound source not only from specified direction but also within defined distance. Specifically, our method employs an improved delay-and-sum technique to obtain directional cues, substantially enhancing the signal from the target direction. We further enhance separation by incorporating the direct-to-reverberant ratio into the input features, enabling the model to better discriminate sources within and beyond a specified distance. Experimental results demonstrate that our proposed method leads to substantial gains across multiple objective metrics. Furthermore, our method achieves state-of-the-art performance on the CHiME-8 MMCSG dataset, which was recorded in real-world conversational scenarios, underscoring its effectiveness for speech separation in practical applications.

SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-Bit Training

The efficiency of attention is important due to its quadratic time complexity. We enhance the efficiency of attention through two key contributions: First, we leverage the new FP4 Tensor Cores in Blackwell GPUs to accelerate attention computation. Our implementation achieves 1038 TOPS on RTX5090, which is a 5x speedup over the fastest FlashAttention on RTX5090. Experiments show that our FP4 attention can accelerate inference of various models in a plug-and-play way. Second, we pioneer low-bit attention to training tasks. Existing low-bit attention works like FlashAttention3 and SageAttention focus only on inference. However, the efficiency of training large models is also important. To explore whether low-bit attention can be effectively applied to training tasks, we design an accurate and efficient 8-bit attention for both forward and backward propagation. Experiments indicate that 8-bit attention achieves lossless performance in fine-tuning tasks but exhibits slower convergence in pretraining tasks. The code will be available at https://github.com/thu-ml/SageAttention.